NVIDIA announced an OpenShell reference architecture combining agent runtime isolation, infrastructure monitoring and hardware enforcement for enterprise and edge deployments.
NVIDIA has announced an OpenShell Reference Architecture designed to help enterprises build, run and supervise AI agents across data centers and edge systems. The announcement, made at GTC Berlin on October 21, brings together software for agent execution with computing and security components intended to contain agent activity and enforce organizational policies.
The architecture combines NVIDIA’s OpenShell secure runtime with Vera Rubin systems for model inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 data-processing units running NVIDIA Sentry. NVIDIA says Sentry can monitor agent behavior outside the host system and enforce policies in hardware, including stopping or quarantining an agent that crosses its configured boundary. Those capabilities are company-described; the announcement does not provide independent evaluations of their effectiveness in production environments.
NVIDIA says the framework can run across major enterprise operating systems and cloud environments, and can extend to edge systems. The company says the architecture is available now. Its partner list spans cloud providers, cybersecurity firms, model developers and enterprise software vendors. SAP, for example, is integrating the architecture with its Joule Studio runtime, while a number of security vendors are working on integrations for agent security products.
The announcement reflects a shift in enterprise AI priorities: as organizations experiment with agents that can use tools and interact with business systems, they need controls over identity, data access and actions—not just model quality. NVIDIA’s approach places those controls across several layers, from the agent’s software runtime to infrastructure-level monitoring. That could appeal to businesses seeking a common deployment framework across different models and environments, but integration with existing security systems and operational policies will determine how useful it proves to be.
The architecture is a vendor-led design, not a universal standard or evidence that autonomous agents are already safe by default. For buyers, the important questions will include how policies are configured, what activity is logged, how exceptions are handled and whether safeguards remain effective across mixed hardware and cloud deployments.